acceptodds
Withdrawn from ICLR 2027

SAFER: Surface-Adaptive Feature Fields for Cross-Source Point Cloud Registration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Point clouds captured by heterogeneous LiDAR sensors exhibit divergent sampling patterns and densities, causing the same physical surfaces to be observed at shifted and non-uniform locations. Existing methods typically encode local geometry with superpoints and their sampled neighborhoods, which can entangle stable surface structure with sensor-dependent sampling variation. Consequently, representations may vary with the sensing mechanism, even when the underlying physical geometry is unchanged. To address this, we propose SAFER, a cross-source point cloud registration framework that lifts superpoint features into a continuous, surface-adaptive feature field by assigning each superpoint an anisotropic Gaussian basis in its local tangent-normal frame. The resulting normalized field affinity provides a geometry-aware prior that encourages compatible support interactions while down-weighting geometrically inconsistent neighborhoods. With this prior, we design Field-Aware Geometric Attention to guide contextual aggregation and introduce a Kernel-Guided Feature Adapter to perform kernel-conditioned descriptor refinement. An overlap-aware field objective further maximizes the consistency of ground-truth-aligned feature fields within their mutually observable regions. To alleviate the scarcity of large-scale cross-source data, we construct two benchmarks. Forest Cross-Source comprises ground-to-aerial LiDAR pairs in unstructured forest scenes, while Indoor Multi-LiDAR covers heterogeneous spinning and solid-state LiDARs in indoor environments. Extensive experiments demonstrate that SAFER improves registration performance across diverse sensor configurations, with notable benefits under severe sampling discrepancies.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-Source Point Cloud Registration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_S62sOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAFER for cross-source point cloud registration. It builds on the UGP pipeline and represents each superpoint with an anisotropic Gaussian support. The support is initialized using local PCA and adjusted through learned scale residuals. Gaussian overlap is then used to guide geometric attention and descriptor refinement. An additional field loss encourages feature consistency in overlapping regions. The authors also construct two registration benchmarks from existing datasets and evaluate performance across several sensor combinations.

Critical strengths

First, the motivation is clear and relevant. Different sensors can sample the same physical surface at different locations and densities. The proposed continuous support provides a reasonable way to reduce dependence on exact superpoint locations.

Second, the unseen sensor-pair experiments show useful improvements. For example, on OS1⇒Horizon, SAFER achieves 56.3% registration recall, compared with 49.7% for the strongest listed baseline, BUFFER-X. The two constructed benchmarks also expand the evaluation of cross-source registration across forest and indoor sensor configurations.

Critical weaknesses

1. The method appears to combine established ideas, while the additional value of its central design is not sufficiently isolated.

SAFER combines PCA-based anisotropic Gaussian supports, Gaussian-overlap affinity, attention bias, residual feature refinement, and a field consistency loss within the existing UGP pipeline. The paper itself discusses earlier probabilistic and continuous registration methods, including anisotropic kernels. Equation (4) also uses the standard Bhattacharyya coefficient between Gaussian distributions.

Using these ideas within cross-source descriptor learning may be useful. However, the current evidence does not clearly establish a substantial methodological advance beyond this combination. The paper needs stronger comparisons to show why surface-adaptive anisotropic fields are necessary, rather than a simpler geometric smoothing or feature-refinement approach.

Table 4 provides component-removal experiments, but important controls are missing. These include isotropic Gaussian supports, a simple distance-based kernel, and an adapter with similar parameter count but without field affinity. Without these controls, it is difficult to separate the benefits of anisotropy, the specific Gaussian affinity, and the added model capacity.

The component gains also require stronger validation. Removing FAGA decreases recall from 89.0% to 88.8%, while removing KGFA decreases it to 88.0%. These results suggest different levels of contribution, but no variability estimates are reported for these differences. The cross-sensor experiments support improved registration performance. However, changes in sampling pattern, density, noise, and field of view occur together. They therefore do not isolate the claimed mechanism of reduced sensitivity to sampling shifts. A controlled experiment that changes sampling while keeping the underlying surface fixed would provide more direct evidence.

2. The performance advantage is limited in several main settings, and recent comparisons and evaluation details remain incomplete.

Some main-table gains over the strongest listed baselines are small. On KITTI Cross-Source, recall increases from 97.3% for SPEAL to 97.6%. On OS0-to-Avia, it increases from 88.4% for GeoTransformer to 89.0%. On OS0-to-Horizon, it increases from 67.7% for UGP to 69.1%. The larger unseen-pair improvements are a positive result, but the paper does not report uncertainty estimates that establish the stability of the smaller gains.

A relevant missing 2026 method is “DFF-Matcher: Robust cross-source registration with density-fused feature and bidirectional consensus matching,” by Rong Guo, Zhenxuan Zeng, Jiang Wu, Xiyu Zhang, Siwen Quan, Zhongwen Hu, Yu Zhu, and Jiaqi Yang, published in the Journal of Visual Communication and Image Representation, 2026. This method directly addresses cross-source density differences and correspondence estimation.

DFF-Matcher reports 98.2% recall, 1.349° rotation error, and 0.312 m translation error on its KITTI-CrossSource evaluation. SAFER reports 97.6%, 0.94°, and 0.48 m, respectively. These published numbers do not establish a direct ranking because identical data splits, preprocessing, and training conditions have not been verified. However, they show why a comparison under a common protocol is important before claiming an advantage over current methods.

There is also an inconsistency in the evaluation description. Section 4.1 states that rotation and translation errors are averaged only over successfully registered pairs. On KITTI, successful registration requires rotation error below 5° and translation error below 2 m. However, Table 2 reports average errors of 5.06° and 2.59 m for Predator. These averages cannot exceed the thresholds if they are computed only over successful pairs. The authors should clarify whether these values use a different averaging rule or are taken from another evaluation protocol.

Questions for authors
  1. What is the main new contribution beyond existing anisotropic continuous kernels and geometric feature refinement? Can the authors provide isotropic-kernel, distance-kernel, and similar-capacity adapter controls?

  2. Can the authors directly measure descriptor consistency and registration performance under controlled sampling changes of the same underlying surface?

  3. Can SAFER and DFF-Matcher be evaluated using the same KITTI-CrossSource split, preprocessing, training conditions, and metric definitions?

  4. How are the Predator errors in Table 2 computed? Please explain their inconsistency with the stated successful-pair averaging rule and report uncertainty estimates for the small recall improvements.

Reviewer_1MSwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SAFER tackles rigid registration of point clouds captured by different LiDAR sensors. It extends the UGP coarse-to-fine pipeline by assigning each superpoint an anisotropic Gaussian support initialized from local PCA and adjusted by learned scale residuals. Gaussian overlap biases self-attention (FAGA) and feature refinement (KGFA), while an overlap-aware field loss encourages consistency under the ground-truth alignment. The authors introduce two cross-source benchmarks and test on KITTI Cross-Source and unseen sensor-pair transfers. Against UGP, registration recall rises from 86.8% to 89.0% on OS0-to-Avia and from 47.1% to 56.3% on the OS1-to-Horizon transfer task.

Critical strengths
  1. The problem is well motivated. Different LiDARs can observe the same surface at different sample locations, and the PCA-aligned Gaussian support is a reasonable way to make coarse features less tied to exact point positions (Sec. 3.1). The overlap measure has a closed form and fits naturally into the matching pipeline.
  2. The unseen sensor-pair tests are useful. In Table 3, SAFER improves registration recall over UGP from 47.1% to 56.3% on OS1-to-Horizon and from 72.8% to 80.2% on Velo-to-Avia, without target-pair fine-tuning. These are the strongest results in the paper.
Critical weaknesses
  1. The ablation studies do not fully substantiate the necessity of the proposed anisotropic field representation. Although Table 4 evaluates the contributions of residual regression, FAGA, and KGFA, it lacks direct comparisons with simpler alternatives, such as isotropic Gaussian supports and distance-based kernels. These comparisons are important for determining whether the performance gains specifically arise from the proposed anisotropic design. In addition, the field loss is evaluated only as an addition to the baseline, without examining its independent contribution within the complete framework. The ablations are also limited to OS0-to-Avia, making it difficult to assess whether the proposed components consistently improve performance across different sensor configurations.
  2. The experimental evaluation lacks sufficient analysis of performance stability and computational efficiency. Although SAFER improves registration recall over UGP on several benchmarks, the magnitude of improvement varies across sensor configurations (Tables 1–3). The paper does not report results across multiple random seeds or provide uncertainty estimates, making it difficult to assess the stability of the reported gains. Furthermore, the proposed method introduces additional Gaussian affinity computations and feature refinement, but their computational overhead is not evaluated. End-to-end inference time comparisons and a more detailed analysis of performance stability would strengthen the experimental evidence.
Questions for authors
  1. Could the authors provide more comprehensive ablation experiments comparing anisotropic Gaussian supports with isotropic Gaussian supports and simpler distance-based kernels? In addition, could the contributions of FAGA, KGFA, and the field loss be further evaluated through combined ablations across different sensor configurations?
  2. Could the authors report the end-to-end inference time of SAFER and all baseline methods in Tables 1–2 under the same hardware and evaluation settings? Additionally, could they provide the mean registration recall and standard deviation across multiple runs with different random seeds to assess the stability of the reported improvements?
Reviewer_N5KdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the sampling-density discrepancies and positional shifts that arise when the same physical surfaces are observed by heterogeneous LiDAR sensors. The authors assign each superpoint an anisotropic Gaussian support initialized from local PCA, and leverage the resulting closed-form field affinity to improve the attention mechanism, the descriptor update, and the training objective. The idea is intuitive, and the closed-form Gaussian overlap makes the design easy to integrate into existing coarse-to-fine registration pipelines. Experiments cover forest, indoor multi-LiDAR, and KITTI scenarios.

Critical strengths

This paper accurately identifies the core limitation of conventional coarse‑to‑fine registration pipelines built upon discrete superpoints. Such methods entangle the intrinsic surface geometry with sensor‑induced sampling variations. Even when the underlying physical geometry remains unchanged, divergent sampling patterns from different LiDAR sensors can lead to inconsistent descriptors, highlighting the key pain point of cross‑source registration. It also distinguishes the sampling mechanism differences between spinning mechanical LiDARs and solid‑state LiDARs. With cumulative‑distribution‑function (CDF) statistics and visualizations, the authors intuitively demonstrate that cross‑source point clouds still suffer from notable sampling shifts even after ground‑truth alignment, which justifies the necessity of this research.

Critical weaknesses

The motivation illustrated in Fig. 1 is not sufficiently clear. As the paper itself notes, cross-source point clouds exhibit density discrepancies, distributional differences, and varying noise levels simultaneously. A CDF of nearest-neighbor distances alone cannot disentangle these confounding factors, and thus does not convincingly isolate the sampling-shift effect that the proposed method is designed to address. I would encourage the authors to enrich this analysis so that the influence of positional shifts can be separated from that of density and noise. Table 1 lacks a comparison on the Inlier Ratio (IR). The central claim of this paper is that the proposed design improves feature consistency and thereby yields more accurate correspondences. IR directly measures the quality of these intermediate correspondences and is therefore essential for validating the claimed mechanism. The IR results of the baseline methods should be added to Table 1. Comparisons with cross-source registration methods are missing. The baselines mainly compared in the paper are same-source registration methods, whereas directly relevant cross-source approaches — such as Cross-PCR, SPEAL, and VRHCF — are not included in the comparison. Since cross-source registration is exactly the target setting of this paper, comparisons with these methods are necessary to substantiate the claimed superiority. The ablation results of the ACFF module itself appear to be missing from the ablation experiments. Given that ACFF is the core component of the proposed method, its standalone contribution needs to be explicitly quantified. Implementation details need to be further supplemented. Several key hyperparameters and computational details are missing, including the number of neighbors k, the scale clipping bounds, the residual strength, the field-affinity injection weight gamma, the weights of each loss term, and the overlap radius r. It should also be clarified whether the attention and the field loss are computed densely over all superpoints or via neighborhood sampling. In addition, the runtime and computational overhead should be reported. Minor Issues Duplicated reference text such as "Eq. quation 2" appears at lines 215 and 262, and there are grammatical errors such as "three challenging benchmarks. demonstrate" in line 101. A careful proofreading pass is needed.

Questions for authors

Please see the weakness part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.